Semantic search and RAG systems with Gemini embeddings (768-3072 dimensions, 8 task types, Cloudflare Vectorize integration).

• Supports flexible output dimensions (768, 1536, 3072, or custom 128-3071) with Matryoshka learning; only 3072-dimensional embeddings are pre-normalized, others require manual normalization for accurate similarity calculations
• Eight task types optimize embeddings for specific use cases: RETRIEVAL_QUERY/DOCUMENT for RAG, SEMANTIC_SIMILARITY for search, CLUSTERING for grouping, CODE_RETRIEVAL_QUERY, QUESTION_ANSWERING, FACT_VERIFICATION, and CLASSIFICATION
• Batch API processes multiple texts in one request but has documented issues: ordering corruption at >500 texts, memory limits at >10k embeddings, and rate limit anomalies; recommended batch size is <100 texts
• Integrates with Cloudflare Vectorize for vector storage and includes complete RAG patterns combining query embedding, semantic search, and LLM generation; free tier allows 100 RPM with exponential backoff required for rate limiting